Component Network Meta-Analysis (CNMA)

Author

Xiaoge Zhang

Published

June 5, 2026

Disentangling the Effects of Combination Therapies

1 Executive Summary: The Rise of Combination Therapies

In modern oncology and chronic disease management, treatments are increasingly formulated as combination therapies. For example, treating advanced melanoma might involve a base chemotherapy (Drug A) combined with a novel immunotherapy (Drug B), creating the combination arm (Drug A + Drug B).

Standard Network Meta-Analysis (NMA) treats every unique combination as a completely separate, independent node in the network. This approach fails to leverage the shared information: the effect of “Drug A” is present in both the monotherapy arm (A) and the combination arm (A + B).

Component Network Meta-Analysis (CNMA) solves this by decomposing treatments into their constituent parts. Instead of estimating the relative effect of “Treatment X vs Treatment Y”, we directly estimate the main effect of “Component A”, “Component B”, etc.

The Reality of Combination Trials A common misconception is that pharmaceutical companies rarely test their drugs in combination with others to avoid confounding the effect of their own drug. In reality, combination trials are the absolute norm in areas like Oncology, HIV, and Hepatitis C. Companies frequently run trials of their new compound on top of an established background therapy or another company’s drug (via collaboration or as Standard of Care) to prove superiority over the existing care standard.

2 The Additive Assumption in CNMA

The most fundamental form of CNMA relies on the Additive Assumption. This assumes that the effect of a combination therapy is exactly equal to the sum of the independent effects of its components.

If \(\theta_A\) is the effect of Drug A, and \(\theta_B\) is the effect of Drug B, then the effect of the combination \(AB\) is modeled as: \[\theta_{AB} = \theta_A + \theta_B\]

2.1 The HTA Perspective: Will NICE Accept This?

In practice, HTA agencies like NICE are highly skeptical of the strict additive assumption, particularly in oncology where drug combinations (e.g., IO + Chemo) are explicitly designed to be synergistic rather than merely additive.

If you submit a pure additive CNMA to NICE, the ERG (Evidence Review Group) will likely challenge it. The accepted practices are:

  1. Interaction Models: Introduce an interaction term (\(\theta_{AB} = \theta_A + \theta_B + \theta_{A*B}\)). If the interaction term is significant, it captures synergy or antagonism. However, robustly estimating interaction terms requires a dense network with trials directly comparing the combination against its individual components (e.g., A+B vs A).
  2. Standard NMA as Base Case: When data is too sparse to estimate interactions, NICE generally prefers treating the combination “A+B” as a completely distinct, independent treatment node (Standard NMA). This avoids making unverified additive assumptions, even if it means sacrificing some statistical power by not pooling the component effects.

2.2 Unique Real-World Scenarios for CNMA

Beyond simply increasing sample size or precision, CNMA is uniquely required in specific scenarios where Standard NMA fails:

  1. Predicting the Efficacy of Unobserved Combinations: Suppose we have trials for Drug A vs Placebo and Drug B vs Placebo. A manufacturer wants to launch a combination therapy Drug A + Drug B but has not yet run a trial for the combination. CNMA allows the prediction of the combination’s efficacy (under the additive assumption) to support early HTA submissions or pricing strategies.
  2. Connecting Disconnected Networks: In Standard NMA, treatments are distinct nodes. If Study 1 compares Drug A vs Drug B and Study 2 compares Drug C vs Drug C + Drug D, these networks are disconnected in standard NMA. You cannot compare A or B with D. CNMA breaks treatments down into components. If Drug A and Drug C share a common component, CNMA can bridge the network at the component level.
  3. Disentangling Heterogeneous “Standard of Care” (SoC) Baselines: In many disease areas, the “control” arm is the investigator’s choice of SoC, which varies across trials (e.g., SoC in Trial 1 contains components X and Y; SoC in Trial 2 contains components X and Z). Standard NMA treats these as different nodes or ignores the difference. CNMA can adjust for the specific components present in the background therapy of each trial.

2.3 Verification of the Additive Assumption

In HTA audits, the additive assumption is the most scrutinized part of a CNMA. It must be verified using both clinical arguments and statistical tests:

  1. The Interaction Model:
    • To test the additive assumption, we can fit an expanded model that includes an interaction term for the combination: \(\theta_{AB} = \theta_A + \theta_B + \theta_{A*B}\).
    • If the 95% Credible Interval (CrI) of the interaction term \(\theta_{A*B}\) crosses 0, it suggests that the data does not provide strong evidence against the additive assumption.
  2. Model Comparison (DIC/WAIC):
    • Compare the Deviance Information Criterion (DIC) or WAIC of the pure additive model with the interaction model.
    • If adding the interaction term does not lead to a meaningful reduction in DIC (e.g., by more than 3 to 5 points), the principle of parsimony favors the simpler pure additive model.

2.4 Boundaries of Applicability

Component NMA is a powerful tool, but it is not universally applicable. Its boundaries are defined by:

  • When to Use (Green Flags):
    • Pharmacological Independence: The components have different mechanisms of action and are not expected to interact biologically.
    • Sparse Networks: When treating combinations as independent nodes results in a disconnected or extremely sparse network, and CNMA is the only way to facilitate indirect comparisons.
  • When NOT to Use (Red Flags):
    • Known Synergy or Antagonism: In areas like oncology where combinations (e.g., IO + Chemo) are explicitly designed to be synergistic, a pure additive model is clinically invalid and will be rejected by agencies like NICE.
    • Data Inseparability: If the network lacks trials comparing the combination directly with its components (e.g., missing trials of A+B vs A or A+B vs B), the interaction term cannot be mathematically identified. Running CNMA in such “thin” networks relies purely on the unverified additive assumption. ## The Minimal Mock Case: Advanced Melanoma Combination Therapy {#3-design-matrix}

To make the Component NMA (CNMA) operation clear, let’s ground it in a concrete clinical scenario answering the specific workflow questions.

2.5 The Scenario & Treatments

We are analyzing clinical trials for Advanced Melanoma. The network evaluates 3 distinct components:

  • Component 1: Placebo (Reference, effect set to 0).
  • Component 2: Drug A (a novel immunotherapy).
  • Component 3: Drug B (another novel immunotherapy).

Across the evidence network, trials do not just test single drugs; they test combinations. We have a total of 4 unique treatments in the network:

  • Trt 1: Placebo.
  • Trt 2: Drug A monotherapy.
  • Trt 3: Drug B monotherapy.
  • Trt 4: Drug A + Drug B combination.

2.6 Extracted Data

Instead of individual patient data, we extract aggregate effect sizes (e.g., Log-Hazard Ratios) from published papers:

  • Study 1 (Trt 1 vs 2): Observed effect = -0.5 (SE = 0.1).
  • Study 2 (Trt 1 vs 3): Observed effect = -0.3 (SE = 0.1).
  • Study 3 (Trt 2 vs 4): Observed effect = -0.3 (SE = 0.1).

2.7 Regression Input Data

We do not treat the 4 treatments as independent categories. We break them down into a Design Matrix \(X\) based on their components:

Treatment Description \(X_2\) (Drug A) \(X_3\) (Drug B)
1 Placebo 0 0
2 Drug A 1 0
3 Drug B 0 1
4 Drug A + B 1 1

The data fed into the Stan regression model consists of: - y: The vector of observed effects from the trials (e.g., [-0.5, -0.3, -0.3]). - X: The rows of the design matrix corresponding to the treatments compared in those trials. - study: Identifiers mapping each arm to its specific study (to handle study-specific baselines \(\mu_j\)).

The linear predictor for Treatment \(k\) in Study \(j\) becomes: \[\theta_{jk} = \mu_j + X_{k,2}\delta_2 + X_{k,3}\delta_3\]

Where: - \(X_{k,2}\) and \(X_{k,3}\) are indicator variables (values 0 or 1) from the Design Matrix. - \(\delta_2\) and \(\delta_3\) are the true underlying marginal effects of Component 2 and Component 3.

2.8 Results

The regression outputs the independent, marginal effects of the components: - \(\delta_2\): The effect of Drug A (estimated to be around -0.5). - \(\delta_3\): The effect of Drug B (estimated to be around -0.3).

With these, we can predict the effect of the combination (Trt 4) relative to Placebo as \(\delta_2 + \delta_3 = -0.8\), even if no trial directly compared Trt 1 vs Trt 4.

3 The Stan Model for CNMA

We implement a standard Normal likelihood (Continuous Data) CNMA model in Stan.

data {
  int<lower=1> N;           // Total number of arms
  real y[N];                // Observed effect size in arm (e.g., mean)
  real<lower=0> se[N];      // Standard error of effect size
  int<lower=1> NS;          // Number of studies
  int<lower=1> study[N];    // Study ID for each arm
  int<lower=1> NC;          // Number of active components
  matrix[N, NC] X;          // Design matrix for components
}
parameters {
  real mu[NS];              // Study-specific baselines
  vector[NC] d;             // Main effects of each component
}
transformed parameters {
  vector[N] theta;
  for (i in 1:N) {
    theta[i] = mu[study[i]] + dot_product(row(X, i), d);
  }
}
model {
  // Priors
  mu ~ normal(0, 5);
  d ~ normal(0, 5);
  
  // Likelihood
  for (i in 1:N) {
    y[i] ~ normal(theta[i], se[i]);
  }
}

4 Simulating and Executing CNMA in R

We simulate a mock network where Drug A has a true effect of -0.5, Drug B has a true effect of -0.3, and their combination is exactly additive (-0.8).

View the code 
library(rstan)

# 1. Define the Design Matrix (4 treatments, 2 active components)
# Col 1: Drug A, Col 2: Drug B
X_matrix <- matrix(c(
    0, 0, # Trt 1: Placebo
    1, 0, # Trt 2: Drug A
    0, 1, # Trt 3: Drug B
    1, 1 # Trt 4: Drug A + Drug B
), ncol = 2, byrow = TRUE)

# 2. Simulate Study Data (Continuous outcome)
set.seed(123)
# Study 1: Trt 1 vs Trt 2
# Study 2: Trt 1 vs Trt 3
# Study 3: Trt 2 vs Trt 4
study_vec <- c(1, 1, 2, 2, 3, 3)
trt_vec <- c(1, 2, 1, 3, 2, 4)

# True Component Effects
d_true <- c(-0.5, -0.3)
mu_true <- c(0.1, 0.2, 0.15) # Study baselines

# Generate observed outcomes
N_arms <- length(study_vec)
X_data <- matrix(0, nrow = N_arms, ncol = 2)
y_obs <- numeric(N_arms) # Pre-allocate a numeric vector of length N_arms initialized with zeros
se_obs <- rep(0.1, N_arms)

for (i in 1:N_arms) {
    X_data[i, ] <- X_matrix[trt_vec[i], ]
    true_theta <- mu_true[study_vec[i]] + sum(X_data[i, ] * d_true)
    y_obs[i] <- rnorm(1, mean = true_theta, sd = se_obs[i])
}

# 3. Stan Data List
stan_data <- list(
    N = N_arms,
    y = y_obs,
    se = se_obs,
    NS = 3,
    study = study_vec,
    NC = 2,
    X = X_data
)

# 4. Run Stan
fit_cnma <- stan(
    model_code = "
data {
  int<lower=1> N;
  real y[N];
  real<lower=0> se[N];
  int<lower=1> NS;
  int<lower=1> study[N];
  int<lower=1> NC;
  matrix[N, NC] X;
}
parameters {
  real mu[NS];
  vector[NC] d;
}
transformed parameters {
  vector[N] theta;
  for (i in 1:N) {
    theta[i] = mu[study[i]] + dot_product(row(X, i), d);
  }
}
model {
  mu ~ normal(0, 5);
  d ~ normal(0, 5);
  for (i in 1:N) {
    y[i] ~ normal(theta[i], se[i]);
  }
}
  ",
    data = stan_data,
    chains = 3,
    iter = 2000,
    warmup = 1000,
    refresh = 0
)

print(fit_cnma, pars = "d")
Inference for Stan model: anon_model.
3 chains, each with iter=2000; warmup=1000; thin=1; 
post-warmup draws per chain=1000, total post-warmup draws=3000.

      mean se_mean   sd  2.5%   25%   50%   75% 97.5% n_eff Rhat
d[1] -0.48       0 0.14 -0.74 -0.57 -0.48 -0.39 -0.19  1286    1
d[2] -0.30       0 0.10 -0.49 -0.36 -0.30 -0.23 -0.10  1626    1

Samples were drawn using NUTS(diag_e) at Mon May 11 22:25:17 2026.
For each parameter, n_eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor on split chains (at 
convergence, Rhat=1).

Based on the Component NMA model, the estimated independent effects of the components are:

  • Component A Effect: -0.48 (True: -0.5)
  • Component B Effect: -0.3 (True: -0.3)

By successfully decomposing the treatments, we can now predict the effect of the combination therapy even if it was never directly compared to placebo in a single trial.

Guide Authored by: Antigravity (Advanced Agentic Coding)